How does a deep research agent differ from RAG?
Standard RAG retrieves once per query and generates from what comes back; the pipeline is fixed. A deep research agent plans first, breaking the question into subgoals, then loops: issue queries, read results, notice gaps, reformulate, search again, run code over findings when useful, then synthesize a cited report. Retrieval is adaptive rather than predetermined, and the loop runs many turns, trading seconds of latency for depth. Agentic RAG sits between the two, adding retrieve evaluate retry behaviour without the full synthesis goal.
Where do deep research reports still need checking?
Known failure modes concentrate at synthesis. An agent can flatten conflicting sources into one confident paragraph without flagging the disagreement, cite a page that does not support the claim, or weight a rumour like a primary source; confidence calibration is weak across these systems, so fluency signals nothing about accuracy. Production setups treat the report as a draft: spot check citations, require primary sources for load bearing claims, and gate automated pipelines with evaluation, often LLM as a judge scored against human labels. Architecturally it overlaps with deep agents: planning, subagents, and a file workspace pointed at research shaped work. Treat every citation as a claim to verify, not evidence in itself.